Improved bgru-based intelligent fault diagnosis method and system for dry-type transformers

Abstract

The application discloses an improved BGRU-based dry-type transformer fault intelligent diagnosis method and system, which is suitable for an online monitoring scene of a cooling fan, continuously collects a load current sequence and a winding temperature sequence, and obtains a cooling fan start-stop state. The method determines an effective thermal time constant based on the winding temperature sequence and in combination with the cooling fan start-stop state, and determines a thermal response blind area based on the effective thermal time constant and a preset proportion coefficient, wherein the preset proportion coefficient is selected according to a load current transient mutation direction. The thermal response blind area is used to constrain the improved BGRU model to handle the confusion of the overload state and the winding short-circuit fault, so that the two types of faults do not occur in the early diagnosis stage in the unidentifiable stage, thereby reducing false positives and keeping the diagnosis output available.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and fault diagnosis technology, and more specifically, to a method and system for intelligent fault diagnosis of dry-type transformers based on an improved BGRU. Background Technology

[0002] Power equipment condition monitoring and fault diagnosis technology is used in the operation of dry-type transformers equipped with cooling fans. It combines an online monitoring system to continuously collect operating parameters such as load current and winding temperature, as well as the start / stop status of the cooling fans, to assess the operating status and determine faults. Existing solutions also include implementation paths that classify and infer the above observations based on time-series models and output fault probabilities or alarms.

[0003] The existing technology has the following shortcomings:

[0004] Existing technologies generally suffer from two problems. Firstly, when the load current experiences a step or transient change, the winding temperature rise lags behind the load change, and a discernible response to the temperature rise may not yet be formed within a short timeframe. Judgments based on short-term observations or single-probability outputs are prone to premature diagnosis and false alarms. Secondly, when the cooling fan's start / stop state changes, the timescale of the temperature rise changes with the cooling state, and the correspondence between temperature rise and fault cause within the same observation window becomes unstable, making it easier to confuse overload conditions with winding short-circuit faults. Therefore, under conditions of superimposed sudden changes and switching, without a unified load-bearing method for the undiscernible stage and a separate diagnostic output after exiting this stage, the diagnostic output is prone to unstable jumps and struggles to complete correction and closed-loop output during the identifiable stage. To address these problems, this invention proposes a solution. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a method and system for intelligent fault diagnosis of dry-type transformers based on an improved BGRU, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for intelligent fault diagnosis of dry-type transformers based on an improved BGRU includes:

[0009] S101 acquires the real-time operating time series data of the dry-type transformer, which includes at least the load current sequence and the winding temperature sequence, and acquires the start-stop status of the cooling fan, and acquires the thermal time constant used to characterize the dynamic response of the winding temperature rise; determines the cooling correction coefficient based on the start-stop status of the cooling fan, and calculates the effective thermal time constant based on the thermal time constant and the cooling correction coefficient.

[0010] When S102 detects a step or transient change in the load current, it determines the time of the change and, starting from the time of the change, determines the end time of the thermal response blind zone based on the effective thermal time constant and the preset proportional coefficient, and generates a thermal response blind zone marker sequence.

[0011] S103 inputs time series data into the improved BGRU model, outputs the original fault classification probability, and performs diagnostic control on the original fault classification probability based on the thermal response blind zone marker sequence. Within the thermal response blind zone, when the maximum probability of the original fault classification probability corresponds to an overload state or a winding short-circuit fault, it outputs a pending state and generates and locks the fault candidate cluster structure. When the thermal response blind zone marker sequence indicates that it has exited the thermal response blind zone and the fault candidate cluster structure is in a locked state, it performs separation diagnosis on the overload state and winding short-circuit fault in the fault candidate cluster structure based on the changing trend of the temperature rise deviation sequence and outputs a diagnostic conclusion.

[0012] In a preferred embodiment, the time series data includes sampling timestamps and is a set of sampled value sequences arranged according to the sampling timestamps. It includes at least a load current sequence and a winding temperature sequence, wherein the load current sequence is a sequence of load current sampled values ​​recorded with each sampling timestamp, and the winding temperature sequence is a sequence of winding measurement point temperature sampled values ​​recorded with each sampling timestamp, and optionally includes an ambient temperature sequence. The sampling period is the time interval between adjacent sampling timestamps. The cooling fan start / stop status is a state sequence aligned point-by-point with the sampling timestamps, and its value indicates whether the cooling fan is on or off. The rated current is the preset rated current of the dry-type transformer. The improved BGRU model uses a dual-channel input sequence composed of the load current sequence and the winding temperature sequence as input. The original fault classification probability is a set of probabilities output by the improved BGRU model for multiple fault types. The operating state characteristics include at least the current load rate and the winding temperature change rate at the moment of abrupt change, wherein the current load rate is determined by the load current and rated current at the moment of abrupt change, and the winding temperature change rate is determined by the difference in winding temperature between adjacent sampling times and the sampling period.

[0013] In a preferred embodiment, the cooling correction factor satisfies the condition that the cooling correction factor when the cooling fan is on is less than the cooling correction factor when the cooling fan is off.

[0014] In a preferred embodiment, detecting a step or transient change in load current includes calculating the magnitude of the load current change rate based on the sampling period and comparing the magnitude with a change threshold. The change threshold is obtained by performing a monotonically decreasing mapping of the current load rate to a preset base threshold, and the change threshold decreases as the current load rate increases. The preset base threshold is a preset load current change rate reference threshold. The monotonically decreasing mapping includes associating the preset base threshold with a mapping coefficient that decreases as the current load rate increases to obtain the change threshold.

[0015] In a preferred embodiment, the preset proportional coefficient is selected as either an increase-load proportional coefficient or a decrease-load proportional coefficient based on the direction of the transient change in load current, with the decrease-load proportional coefficient being greater than the increase-load proportional coefficient. The preset proportional coefficient is determined by the operating state characteristics, including applying a preset gain coefficient to the proportional coefficient after direction selection based on the current load rate and the winding temperature change rate, so that the preset proportional coefficient is larger when the current load rate or the winding temperature change rate is larger, thereby delaying the end time of the thermal response blind zone accordingly. The preset proportional coefficient is used to map the effective thermal time constant to the duration of the thermal response blind zone.

[0016] In a preferred embodiment, the thermal response blind zone marker sequence is aligned point-by-point with the time series data; and when the start-stop state of the cooling fan changes, the cooling correction coefficient is updated and the effective thermal time constant is updated synchronously. Based on the updated effective thermal time constant, the end time of the thermal response blind zone is recalculated and updated, and the thermal response blind zone marker sequence is updated synchronously. The recalculation and update are performed only on the remaining blind zone portion after the change in the start-stop state of the cooling fan, so that the updated end time of the thermal response blind zone does not fall into the past time interval. The blind zone marker in the thermal response blind zone marker sequence is used to indicate the time between the abrupt change and the end time of the blind zone, and the non-blind zone marker is used to indicate the time outside the blind zone.

[0017] In a preferred embodiment, the fault candidate cluster structure includes the following fields: a candidate hypothesis set field, used to record candidate fault types, including at least overload conditions and winding short-circuit faults; a lock flag field, used to indicate whether the current state is locked; a lock start time field, used to record the start time of entering the locked state; a mutation time field, used to record the mutation time that triggers the thermal response blind zone calculation; a thermal response blind zone end time field, used to record the end time of the thermal response blind zone; a pending state field, used to output a pending state during the lockout period and to indicate that a single diagnostic conclusion will not be output for the time being; and a deviation history queue field, used to carry recent sequence segments of the temperature rise deviation sequence to support split determination.

[0018] In a preferred embodiment, the reference temperature rise trajectory is generated by a lightweight temperature rise following model. The temperature rise following model generates a reference temperature rise trajectory corresponding to load changes based on real-time load current, and is corrected by combining ambient temperature and initial temperature conditions. The temperature rise following speed is limited by an effective thermal time constant. The temperature rise deviation sequence is the difference between the winding temperature and the reference temperature rise trajectory, and the recent sequence segments of the temperature rise deviation sequence are written into the deviation history queue field.

[0019] In a preferred embodiment, after exiting the thermal response blind zone, the separation diagnosis of overload state and winding short circuit fault includes judging the changing trend of temperature rise deviation sequence after the load current enters a stable or declining state. If the temperature rise deviation sequence decreases or tends to stabilize, it is judged as an overload state. If the temperature rise deviation sequence continues to increase, it is judged as a winding short circuit fault and a diagnosis alarm is triggered. The load current entering a stable or declining state includes the magnitude of the load current change rate being less than a preset stable threshold within a preset number of consecutive samplings or the load current showing a continuous downward trend.

[0020] The intelligent fault diagnosis system for dry-type transformers based on an improved BGRU includes: a data acquisition and effective thermal time constant generation unit, used to acquire time series data, cooling fan start / stop status, and thermal time constant; determine the cooling correction coefficient based on the cooling fan start / stop status; and calculate the effective thermal time constant based on the thermal time constant and the cooling correction coefficient; a sudden change triggering and blind zone boundary calculation and marking unit, used to detect step or transient changes in load current and determine the moment of the change; using the moment of the change as the starting point, determine the end time of the thermal response blind zone based on the effective thermal time constant and a preset proportional coefficient; and generate a thermal response blind zone marking sequence aligned with the time series data; and a model inference and diagnostic control unit, used to input the time series data into the improved BGRU model and output the original fault classification probability. Based on the thermal response blind zone marking sequence, diagnostic control is performed on the original fault classification probability. Within the thermal response blind zone, when the maximum probability of the original fault classification probability corresponds to an overload state or a winding short-circuit fault, a pending state is output and a fault candidate cluster structure is generated and locked. The post-blind zone split diagnostic unit is used to split and diagnose the overload state and winding short-circuit fault in the fault candidate cluster structure based on the changing trend of the temperature rise deviation sequence when the thermal response blind zone marking sequence indicates that it has exited the thermal response blind zone and the fault candidate cluster structure is in a locked state, and output a diagnostic conclusion. When a winding short-circuit fault is diagnosed, a diagnostic alarm is triggered. The preset parameter storage area is used to store preset parameters for determining the sudden change trigger and the end time of the thermal response blind zone, which are read by the sudden change trigger and blind zone boundary calculation and marking unit.

[0021] The present invention presents an intelligent fault diagnosis method and system for dry-type transformers based on an improved BGRU, and its effects and advantages are as follows:

[0022] This invention introduces an effective thermal time constant and a thermal response blind zone into the online monitoring and diagnostic process of dry-type transformers, and constructs a diagnostic control and pending output mechanism based on the thermal response blind zone. This ensures that the diagnostic output when the load current experiences a step or transient change is constrained by the temperature rise response time scale. Before a recognizable response is formed during the temperature rise phase, a single diagnostic conclusion is not given for overload conditions and winding short-circuit faults, but a pending state is output instead. This helps reduce false alarms and repetitive conclusions caused by premature diagnosis, thereby improving the stability and consistency of the diagnostic output for this confusing fault. Furthermore, in an optional embodiment, this invention can generate and lock a fault candidate cluster structure within the thermal response blind zone. After exiting the thermal response blind zone, it combines the reference temperature rise trajectory and temperature rise deviation sequence to perform split diagnosis of overload conditions and winding short-circuit faults and output diagnostic conclusions. A diagnostic alarm is triggered when a winding short-circuit fault is diagnosed. This further enhances the usability of diagnostic results under sudden operating conditions without adding additional observations or changing existing online acquisition conditions. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0024] Figure 2 This is a schematic flowchart illustrating the mutation triggering and blind zone boundary calculation and marking of the present invention;

[0025] Figure 3 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] This invention provides an intelligent fault diagnosis method for dry-type transformers based on an improved BGRU, applicable to online monitoring scenarios where cooling fans are configured and load current and winding temperature sequences can be continuously acquired to obtain the start-stop status of the cooling fans. When the load current experiences a step or transient change and the cooling fan switches between start and stop, the temperature rise response lags and changes with the time scale. This causes the diagnostic output, before the temperature rise has formed a recognizable response stage, to be easily confused with overload conditions and winding short-circuit faults, resulting in premature diagnosis.

[0028] Therefore, this invention aims to avoid premature diagnosis during the unidentifiable stage and complete the closed-loop output from pending to confirmed diagnosis during the identifiable stage, under the aforementioned constraints. This invention obtains an effective thermal time constant based on the winding temperature sequence and the start / stop status of the cooling fan. When a sudden change in load current triggers the process, a thermal response blind zone is determined based on the effective thermal time constant and a preset proportional coefficient. The preset proportional coefficient is selected according to the direction of the transient change in load current and is jointly determined by operating state characteristics, which include at least the current load rate and the winding temperature change rate. The thermal response blind zone is used to constrain the improved BGRU model's handling of the confusion between overload conditions and winding short-circuit faults, preventing premature diagnosis of both types of faults during the unidentifiable stage, thereby reducing false alarms and maintaining the availability of diagnostic output.

[0029] Based on the above design, this invention constructs a processing flow consisting of steps S101 to S103 sequentially. (Refer to...) Figure 1 , Figure 1 This is a schematic diagram of the method flow of the present invention, which includes:

[0030] Step S101, data acquisition and effective thermal time constant generation, is used in the upstream preparation stage of this method to acquire online monitoring data and form an equivalent thermal time scale input consistent with the cooling conditions. This ensures that subsequent step S102 can calculate the thermal response blind zone boundary with a consistent caliber when a sudden change in load current is triggered. This step takes time series data, cooling fan start / stop status, and thermal time constant as input, aligns the timestamp caliber, handles missing points, determines the cooling correction coefficient, calculates the effective thermal time constant, and outputs a cooling correction coefficient sequence and an effective thermal time constant sequence. The corresponding values ​​can be read from the nearest timestamp for subsequent blind zone boundary calculation. Among them, time series data is used to provide the time series observation entry of load current sequence and winding temperature sequence and serve as the basis for subsequent model inference input. Cooling fan start-stop status is used to characterize cooling conditions and drive the selection of cooling correction coefficient. Thermal time constant is used to characterize the speed of temperature rise dynamic response and serve as the benchmark quantity for equivalent time scale calculation. Cooling correction coefficient is used to correct thermal time scale to obtain effective thermal time constant, so as to maintain the consistency of blind zone boundary calculation caliber when thermal time scale changes caused by cooling fan start-stop switching.

[0031] Step S102, "Sudden Change Trigger and Dead Zone Boundary Calculation and Marking," is used to determine the moment of change and define the thermal response dead zone boundary when a step or transient change in load current is detected. Simultaneously, it generates a thermal response dead zone marker sequence aligned point-by-point with the time series data, enabling subsequent step S103 to perform diagnostic control from a unified entry point. This step takes the load current sequence, cooling fan start / stop status, effective thermal time constant, preset proportional coefficient, and operating status characteristics as inputs to determine the moment of change and the end time of the thermal response dead zone, and generates a thermal response dead zone marker sequence. When the cooling fan start / stop status changes and the effective thermal time constant is updated, the dead zone end time and marker sequence are updated synchronously for subsequent step S103 to read and use. The thermal response blind zone is the time interval from the moment of abrupt change to the end of the thermal response blind zone. It is used to avoid premature diagnosis of the confusing fault pair of overload state and winding short circuit fault before the temperature rise response has formed a discernible difference. The preset proportional coefficient is used to map the effective thermal time constant to the duration of the blind zone to determine the end of the thermal response blind zone. It is constrained by the direction of the abrupt change and jointly determined by the operating state characteristics. The thermal response blind zone marking sequence is used to indicate whether it is in the thermal response blind zone at each sampling moment and to provide a unified reading caliber for subsequent diagnostic control.

[0032] Step S103, model inference, diagnostic control, and closed-loop decomposition, is used to input time series data into the improved BGRU model based on the generated thermal response blind zone marker sequence, outputting the original fault classification probability. Based on the thermal response blind zone marker sequence, diagnostic control within the blind zone is implemented. When exiting the blind zone but still under locking, the decomposition and diagnostic output of overload and winding short-circuit faults are completed. Specifically, within the thermal response blind zone, when the highest probability of the original fault classification probability corresponds to either an overload or winding short-circuit fault, a pending state is output, and a fault candidate cluster structure R102 is generated and locked. The pending state is used to prevent the output of a single diagnostic conclusion to avoid confusion and premature diagnosis of fault pairs. The fault candidate cluster structure R102 carries candidate hypotheses and the locked state during the locking period, records the abrupt change time and the end time of the thermal response blind zone, and carries a deviation history queue, thus providing a unified reading caliber for evidence generation and decomposition diagnostics after exiting the blind zone.

[0033] The implementation process and operational effects of the method of the present invention will be described in detail below with reference to specific embodiments. It should be understood that the embodiments are only used to illustrate the technical solution of the present invention, and not to limit it. The relevant steps, parameters, and module divisions can be appropriately adjusted without changing the essence of the invention.

[0034] For ease of understanding, the following embodiments are described under a unified system architecture, which can be modified equivalently according to actual needs. In an optional embodiment, the system consists of a data acquisition and effective thermal time constant generation unit, a mutation triggering and blind zone boundary calculation and marking unit, a model inference and diagnostic control unit, a blind zone post-split diagnostic unit, and a preset parameter storage area. (Refer to...) Figure 3 , Figure 3 This is a schematic diagram of the system structure of the present invention.

[0035] Specifically, the online monitoring data, as system input, includes at least dry-type transformer time series data, cooling fan start / stop status, and thermal time constant. The time series data includes at least load current sequence and winding temperature sequence. The data acquisition and effective thermal time constant generation unit receives the online monitoring data, determines the cooling correction coefficient based on the cooling fan start / stop status, calculates the effective thermal time constant based on the thermal time constant and the cooling correction coefficient, and then outputs the cooling correction coefficient sequence and the effective thermal time constant sequence. The preset parameter storage area stores preset basic thresholds, mapping coefficients, load increase ratio coefficients, load decrease ratio coefficients, and preset gain coefficients, which are read by the sudden change triggering and blind zone boundary calculation and marking unit. The sudden change triggering and blind zone boundary calculation and marking unit receives the effective thermal time constant sequence, detects step or transient sudden changes based on the load current sequence in the online monitoring data to determine the sudden change time and the end time of the thermal response blind zone, and generates a thermal response blind zone marking sequence based on the effective thermal time constant and the preset ratio coefficient, and then outputs the sudden change time, the end time of the thermal response blind zone, and the thermal response blind zone marking sequence. The model inference and diagnostic control unit receives the thermal response blind zone marker sequence, abrupt change time, and thermal response blind zone end time. It inputs time-series data from online monitoring into the improved BGRU model to output the original fault classification probability. Within the blind zone, it outputs pending states and generates and locks the fault candidate cluster structure R102. The post-blind zone split diagnostic unit performs split diagnostics on overload conditions and winding short-circuit faults based on R102 after exiting the blind zone, outputs diagnostic conclusions, and triggers a diagnostic alarm when a winding short-circuit fault is diagnosed. The system sends the original fault classification probability, pending states, and split diagnostic output to the diagnostic output terminal.

[0036] In an optional embodiment, step S101 is performed by the data acquisition and effective thermal time constant generation unit, which is used to align the start-stop status of the cooling fan with the time series data under the sampling timestamp reference, and generate an effective thermal time constant on this basis, so that subsequent steps can read the corresponding effective thermal time constant as the input for blind zone boundary calculation at the moment of abrupt change.

[0037] The time-series data includes sampling timestamps and includes at least a load current sequence and a winding temperature sequence, and optionally an ambient temperature sequence. Specifically, the time-series data is a set of sampled value sequences arranged by sampling timestamps; the load current sequence is a sequence of load current sampled values ​​recorded with each sampling timestamp; the winding temperature sequence is a sequence of winding measurement point temperature sampled values ​​recorded with each sampling timestamp; and the sampling period is the time interval between adjacent sampling timestamps.

[0038] This step's processing flow includes timestamp alignment and missing point handling, start / stop state debouncing and cleaning, cooling correction coefficient determination, and effective thermal time constant calculation and output organization. The missing point completion strategy, start / stop state debouncing parameters, and cooling correction coefficient value rules are provided by a preset parameter storage area.

[0039] In timestamp alignment and missing point handling, starting from the time series data, using its sampling timestamp as the reference time axis, missing points in the load current sequence and winding temperature sequence are filled according to a preset filling strategy, resulting in load current sequences and winding temperature sequences corresponding point-by-point on the same timestamp set. Long, continuously missing segments are marked as unusable segments to avoid using unreliable input in subsequent calculations. In start-stop state debounce and cleaning, starting from the start-stop state of the cooling fan, the start-stop state is mapped to each sampling timestamp according to the reference time axis, and the previous valid state is maintained for the missing measurement interval, resulting in a start-stop state sequence aligned point-by-point with the time series data. Then, the start-stop switching time is confirmed according to the debounce and abnormal jump correction rules, and short-duration states are removed, ensuring that the start-stop state only retains two values: on and off, and that the time order is not rearranged. The cooling fan start-stop state is a state sequence aligned point-by-point with the sampling timestamp, and its value represents whether the cooling fan is on or off. In determining the cooling correction coefficient, starting from the start-stop state sequence after debouncing and cleaning, and following the selection rule that the value of the start condition is less than that of the stop condition, the start-stop state at each sampling moment is mapped to the corresponding cooling correction coefficient value. This ensures that the switching time of the cooling correction coefficient is consistent with the confirmed start-stop switching time, thus obtaining a cooling correction coefficient value sequence that is point-by-point associated with the sampling timestamp. In the effective thermal time constant calculation and output organization, starting from the thermal time constant and the cooling correction coefficient value, the effective thermal time constant value is calculated point-by-point according to the equivalent time scale calculation rule. The cooling correction coefficient value, the effective thermal time constant value, and the sampling timestamp are associated and saved, so that subsequent steps can read the corresponding effective thermal time constant as the input for blind zone boundary calculation after determining the abrupt change time. The data acquisition and effective thermal time constant generation unit provides the cooling correction coefficient sequence and the effective thermal time constant sequence to the abrupt change triggering and blind zone boundary calculation and marking unit for reading, and provides the completed load current sequence and the completed winding temperature sequence as a unified input for subsequent model inference to the model inference and diagnostic control unit.

[0040] To ensure the reproducibility of the calculation methods for the cooling correction factor and the effective thermal time constant, the following minimal formula set can be used. Let the sampling timestamp be... Assume the cooling fan is in start / stop state as follows: ,in This indicates that the cooling fan is on. This indicates that the cooling fan is off. Assume the thermal time constant is... This is used to characterize the dynamic response of the winding temperature rise and serves as a benchmark quantity for calculations on an equivalent time scale. Let the value be taken as the fan start-up condition. The value for the fan shutdown condition is... and satisfy Cooling correction factor Mapping point by point according to start / stop status Examples are acceptable. , , used to give the satisfaction This is a set of intuitive values ​​that can be replaced by calibration results in engineering applications. Effective thermal time constant Calculated from thermal time constant and cooling correction factor ;in Sampling timestamp The corresponding effective thermal time constant, Sampling timestamp The corresponding cooling correction factor, is the thermal time constant.

[0041] In an optional embodiment, step S102 is executed by the sudden change triggering and blind zone boundary calculation and marking unit. This unit provides a unified blind zone boundary and point-by-point marking caliber after the load current sudden change is triggered, and synchronously updates the blind zone boundary and marking sequence when the effective thermal time constant changes due to the start-stop switching of the cooling fan, to avoid the blind zone boundary becoming disconnected from the actual thermal response timescale. The preset proportional coefficient is provided by the preset parameter storage area and read by the sudden change triggering and blind zone boundary calculation and marking unit. The preset parameter storage area stores preset base thresholds, mapping coefficients, load increase proportional coefficients, load decrease proportional coefficients, and preset gain coefficients. The operating status characteristics include at least the current load rate and winding temperature change rate at the moment of sudden change. The current load rate is determined by the load current and rated current at the moment of sudden change, and the winding temperature change rate is determined by the winding temperature difference between adjacent sampling times and the sampling period.

[0042] This step's processing flow includes mutation detection and mutation timing determination, blind zone end time calculation, blind zone recalculation and update under changing cooling conditions, and blind zone marker sequence generation and synchronous update. (Refer to...) Figure 2 , Figure 2 A flowchart illustrating the calculation and labeling of mutation triggering and blind zone boundaries.

[0043] In the mutation detection and mutation time determination, starting from the load current sequence, the load current change rate is calculated according to the calculation caliber of the load current change rate. This change rate is compared with the mutation judgment threshold to generate a mutation trigger mark. The sampling time when the trigger condition is first met is taken as the mutation time. The mutation judgment threshold is obtained by performing a monotonically decreasing mapping of the current load rate to a preset base threshold, and the higher the current load rate, the lower the mutation judgment threshold. The preset base threshold is a preset load current change rate benchmark threshold, and the mapping coefficient is a preset mapping coefficient. The preset base threshold is the benchmark trigger threshold for the load current change rate magnitude, and the mapping coefficient is used to adjust the magnitude by which the mutation judgment threshold decreases as the current load rate increases. To suppress false triggering caused by single-point noise, a short window smoothing can be used for the load current change rate, and a minimum hold condition can be set so that the trigger mark remains consistent within a number of consecutive sampling points before the mutation time is confirmed. At the same time, after confirming the mutation, a window for suppressing repeated triggering is set to avoid the same mutation process being triggered repeatedly.

[0044] In calculating the end time of the blind zone, starting from the abrupt change time and direction, the load increase or decrease proportional coefficient is selected according to the direction of the abrupt change. Combined with the operating state characteristics, a preset proportional coefficient is jointly determined so that the blind zone duration increases accordingly when the current load rate or the winding temperature change rate is higher, thus delaying the end time of the thermal response blind zone. Then, starting from the effective thermal time constant, the blind zone duration is mapped according to the preset proportional coefficient, and the end time of the thermal response blind zone is calculated accordingly, ensuring that the blind zone boundary does not decrease with the increase of the effective thermal time constant. The joint determination includes applying a preset gain coefficient to the proportional coefficient selected after direction selection based on the current load rate and the winding temperature change rate.

[0045] In the blind zone recalculation update under changing cooling conditions, when the start-stop state of the cooling fan changes and triggers the update of the effective thermal time constant, the blind zone duration is recalculated based on the abrupt change time and the elapsed duration, and only the remaining blind zone portion is updated. This ensures that the updated end time of the blind zone does not fall into the elapsed time interval, and that the remaining blind zone duration is truncated to zero when it is calculated to be negative, thereby avoiding negative duration and preventing blind zone boundary bounce.

[0046] In the generation and synchronous update of the blind zone marker sequence, starting from the set of sampling timestamps of the time series data, blind zone markers or non-blind zone markers are assigned to each sampling moment according to the abrupt change moment and the end moment of the thermal response blind zone. This results in a thermal response blind zone marker sequence that is point-by-point aligned with the load current sequence and winding temperature sequence. After updating at the end moment of the blind zone, the marker sequence is synchronously recalculated according to the updated blind zone boundary to ensure that the blind zone boundary and the point-by-point markers are consistent under the same caliber. The abrupt change triggering and blind zone boundary calculation and marking unit provides the thermal response blind zone marker sequence, abrupt change moment, and end moment of the thermal response blind zone to subsequent diagnostic control readings.

[0047] To ensure the reproducibility of the generation criteria for the mutation determination threshold, the preset proportional coefficient, and the end time of the thermal response blind zone, the following minimal formula set can be used. Let the mutation time be... Let the sampling period be... Let the load current sequence be... The winding temperature sequence is Let the rated current be... Assume the current load rate is... Assume the winding temperature change rate is... Let the magnitude of the load current change rate be... Let the preset basic threshold be... The mutation detection threshold is Then, optionally adopt ,in The magnitude of the load current change rate The baseline trigger threshold, The mapping coefficient is used to adjust the mutation detection threshold as a function of the current load rate. The magnitude of the increase and decrease, and both are provided by the preset parameter storage area.

[0048] Let the mutation direction determination quantity be... ,when The mutation direction is considered to be increasing, when The direction of the mutation is assumed to be decreasing. Let the load increase ratio coefficient be... The load reduction ratio is and satisfy The load increase ratio coefficient and load decrease ratio coefficient are direction-dependent ratio coefficients that map the effective thermal time constant to the duration of the dead zone, and both are provided by the preset parameter storage area. The base ratio coefficient is selected according to the direction of the abrupt change. Examples are acceptable. , This is used to reflect a more conservative directional constraint on the blind zone boundary under load reduction scenarios.

[0049] Based on this, a gain correction is performed on the preset proportional coefficient according to the operating state characteristics, so that the preset proportional coefficient is larger when the current load rate is higher or the winding temperature change rate is higher. For example .in and These are preset gain coefficients, used to characterize the current load rate term and the winding temperature change rate term relative to the preset proportional coefficient, respectively. The gain contribution is provided by the preset parameter storage area.

[0050] Let the effective thermal time constant corresponding to the abrupt change be . Let the duration of the thermal response blind zone be... The end time of the thermal response blind zone is ,but , When the start / stop state of the cooling fan changes, triggering an update of the effective thermal time constant, let the start / stop switching time be... The updated effective thermal time constant is Let the elapsed time be . To avoid blind zone boundary bounce and only update the remaining blind zone portion, take... , .

[0051] Let the thermal response blind zone labeling sequence be... Then assign values ​​point by point according to the blind zone boundary. ,in This indicates that it is in the thermal response blind zone. This indicates that it is not in the thermal response blind zone.

[0052] In an optional embodiment, step S103 is executed collaboratively by the model inference and diagnostic control unit and the post-blind zone split diagnostic unit. This is achieved by stabilizing the output of pending and locking candidate cluster structures within the blind zone as an entry point, and forming a closed-loop output from pending to confirmed diagnosis in the identifiable stage after exiting the blind zone. The processing flow of this step includes input alignment and model inference, diagnostic control and candidate cluster locking within the blind zone, and evidence generation and split diagnostic diagnosis after exiting the blind zone.

[0053] In input alignment and model inference, based on the supplemented load current sequence and supplemented winding temperature sequence provided in step S101, the input window is constructed and the channel is organized according to the input channel aperture of the improved BGRU model to obtain a dual-channel input sequence, which is then input into the improved BGRU model to obtain the original fault classification probability.

[0054] In the blind zone diagnostic control and candidate cluster locking, starting from the thermal response blind zone marker sequence, it is determined whether the current sampling time is within the thermal response blind zone according to the distinction rules between blind zone markers and non-blind zone markers. When it is within the thermal response blind zone and the highest probability of the original fault classification corresponds to an overload state or a winding short-circuit fault, a pending state is output and a fault candidate cluster structure R102 is generated or updated, causing R102 to enter a locked state and the candidate hypothesis set to be recorded as at least an overload state and a winding short-circuit fault. To ensure that the blind zone boundary field within R102 is verifiable, the abrupt change time and the thermal response blind zone end time are determined based on the start and end positions of the blind zone markers in the thermal response blind zone marker sequence and written into R102. At the same time, the locking start time is recorded and the pending state field is maintained for external output. In an optional implementation, the fault candidate cluster structure R102 includes a candidate hypothesis set field, a lock flag field, a lock start time field, a mutation time field, a thermal response blind zone end time field, a pending state field, and a deviation history queue field. The candidate hypothesis set field records at least overload states and winding short-circuit faults, and the deviation history queue field is used to carry recent sequence segments of the temperature rise deviation sequence to support split determination.

[0055] In the evidence generation and deconstruction confirmation after exiting the blind zone, when the thermal response blind zone marker sequence indicates that the thermal response blind zone has been exited and R102 is in a locked state, a reference temperature rise trajectory is generated from time series data and combined with the effective thermal time constant. The following speed of the temperature rise following model is limited by the effective thermal time constant to ensure that the reference temperature rise trajectory is consistent with the actual cooling conditions on the thermal response time scale. The lightweight temperature rise following model generates a reference temperature rise trajectory corresponding to load changes based on real-time load current, and makes corrections based on ambient temperature and initial temperature conditions when generating the reference temperature rise trajectory. Subsequently, according to the temperature rise deviation sequence calculation caliber, the winding temperature sequence and the reference temperature rise trajectory are subtracted point by point to obtain the temperature rise deviation sequence. At the same time, the most recent segments of the temperature rise deviation sequence are written into the deviation history queue of R102 to form the entry point for deconstruction evidence. Subsequently, starting from the load current sequence, it is determined whether the load current has entered a stable or declining phase according to the rules for judging the stable or declining phase. After entering the stable or declining phase, a split judgment is performed based on the changing trend of the temperature rise deviation sequence. If the temperature rise deviation sequence decreases or tends to stabilize, the pending condition is removed and it is judged as an overload state. If the temperature rise deviation sequence continues to increase, it is judged as a winding short circuit fault and a confirmed alarm is triggered. Optionally, when the load current is continuously at a stable high load and has not entered the declining phase for a long time, if the temperature rise deviation sequence continues to increase and meets the preset duration threshold, a fallback split is performed and output according to the winding short circuit fault caliber and a confirmed alarm is triggered.

[0056] After the above process, this step uses the candidate hypothesis set field and the lock flag field to record the pending lock entry point within the blind zone, uses the mutation time field and the thermal response blind zone end time field to record the blind zone boundary to support the exit from the blind zone determination, uses the deviation history queue field to carry recent evidence fragments of the temperature rise deviation sequence, and updates the pending state field and the lock flag field after the split determination is completed to output the diagnostic conclusion. When a winding short circuit fault is diagnosed, a diagnosis alarm is triggered and short-term repeated triggering is suppressed according to the preset deduplication and retention strategy. The post-blind zone split diagnosis unit sends the split diagnosis output to the diagnostic output terminal. The preset duration threshold and the deduplication and retention strategy are provided by the preset parameter storage area.

[0057] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0058] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and inventive constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0059] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0060] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0061] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent fault diagnosis of dry-type transformers based on an improved BGRU, characterized in that, include: S101 acquires the real-time operating time series data of the dry-type transformer, which includes at least the load current sequence and the winding temperature sequence, and acquires the start-stop status of the cooling fan, and acquires the thermal time constant used to characterize the dynamic response of the winding temperature rise; determines the cooling correction coefficient based on the start-stop status of the cooling fan, and calculates the effective thermal time constant based on the thermal time constant and the cooling correction coefficient. When S102 detects a step or transient change in the load current, it determines the time of the change and, starting from the time of the change, determines the end time of the thermal response blind zone based on the effective thermal time constant and the preset proportional coefficient, and generates a thermal response blind zone marker sequence. S103 inputs time series data into the improved BGRU model, outputs the original fault classification probability, and performs diagnostic control on the original fault classification probability based on the thermal response blind zone label sequence; Within the thermal response blind zone, when the maximum probability of the original fault classification corresponds to an overload state or a winding short-circuit fault, an undetermined state is output and a fault candidate cluster structure is generated and locked. When the thermal response blind zone marker sequence indicates that the thermal response blind zone has been exited and the fault candidate cluster structure is in a locked state, the overload state and winding short circuit fault in the fault candidate cluster structure are separated and diagnosed based on the changing trend of the temperature rise deviation sequence, and the diagnostic conclusion is output.

2. The intelligent fault diagnosis method for dry-type transformers based on improved BGRU according to claim 1, characterized in that, The time series data includes sampling timestamps and is a set of sampled value sequences arranged according to the sampling timestamps. It includes at least a load current sequence and a winding temperature sequence, wherein the load current sequence is a sequence of load current sampled values ​​recorded with each sampling timestamp, and the winding temperature sequence is a sequence of winding measurement point temperature sampled values ​​recorded with each sampling timestamp. The sampling period is the time interval between adjacent sampling timestamps. The cooling fan start / stop status is a state sequence aligned point-by-point with the sampling timestamps, and the values ​​represent whether the cooling fan is on or off. The improved BGRU model uses a dual-channel input sequence composed of the load current sequence and the winding temperature sequence as input. The original fault classification probability is a set of probabilities output by the improved BGRU model for multiple fault types. The operating state characteristics include at least the current load rate and the winding temperature change rate at the moment of abrupt change, wherein the current load rate is determined by the load current and rated current at the moment of abrupt change, and the winding temperature change rate is determined by the difference in winding temperature between adjacent sampling times and the sampling period.

3. The intelligent fault diagnosis method for dry-type transformers based on the improved BGRU according to claim 2, characterized in that, The time series data may include ambient temperature series.

4. The intelligent fault diagnosis method for dry-type transformers based on the improved BGRU according to claim 1, characterized in that, The cooling correction factor satisfies the condition that the cooling correction factor when the cooling fan is on is less than the cooling correction factor when the cooling fan is off.

5. The intelligent fault diagnosis method for dry-type transformers based on the improved BGRU according to claim 4, characterized in that, Detecting a step or transient change in load current involves calculating the magnitude of the load current change rate based on the sampling period and comparing this magnitude with a change threshold. The change threshold is obtained by performing a monotonically decreasing mapping of the current load rate to a preset base threshold, and the change threshold decreases as the current load rate increases. The preset base threshold is a preset load current change rate benchmark threshold. The monotonically decreasing mapping involves associating the preset base threshold with a mapping coefficient that decreases as the current load rate increases to obtain the change threshold.

6. The intelligent fault diagnosis method for dry-type transformers based on the improved BGRU according to claim 5, characterized in that, The preset proportional coefficient is selected based on the direction of the transient change in load current, and the load reduction proportional coefficient is greater than the load increase proportional coefficient. The preset proportional coefficient is determined by the operating state characteristics, including applying a preset gain coefficient to the proportional coefficient after the direction selection based on the current load rate and the winding temperature change rate. This makes the preset proportional coefficient larger when the current load rate or the winding temperature change rate is larger, thereby delaying the end time of the thermal response blind zone accordingly. The preset scaling factor is used to map the effective thermal time constant to the duration of the thermal response blind zone.

7. The intelligent fault diagnosis method for dry-type transformers based on the improved BGRU according to claim 6, characterized in that, The thermal response blind zone marker sequence is aligned point-by-point with the time series data. When the start-stop state of the cooling fan changes, the cooling correction coefficient is updated and the effective thermal time constant is updated synchronously. Based on the updated effective thermal time constant, the end time of the thermal response blind zone is recalculated and updated, and the thermal response blind zone marker sequence is updated synchronously. The recalculation and update are performed only on the remaining blind zone portion after the change in the start-stop state of the cooling fan, so that the updated end time of the thermal response blind zone does not fall into the past time interval. The blind zone marker in the thermal response blind zone marker sequence is used to indicate the time between the abrupt change and the end time of the blind zone, and the non-blind zone marker is used to indicate the time outside the blind zone.

8. The intelligent fault diagnosis method for dry-type transformers based on the improved BGRU according to claim 7, characterized in that, The fault candidate cluster structure includes the following fields: a candidate hypothesis set field, used to record candidate fault types, including at least overload conditions and winding short-circuit faults; and a lock flag field, used to indicate whether the current state is locked. The lock start time field is used to record the start time of entering the locked state; The mutation time field is used to record the mutation time that triggers the calculation of the thermal response blind zone; The thermal response blind zone end time field is used to record the end time of the thermal response blind zone; The pending status field is used to output a pending status during the lockout period and to indicate that a single confirmed diagnosis conclusion will not be output for the time being; the deviation history queue field is used to carry recent sequence segments of the temperature rise deviation sequence to support split judgment.

9. The intelligent fault diagnosis method for dry-type transformers based on the improved BGRU according to claim 8, characterized in that, The reference temperature rise trajectory is generated by a lightweight temperature rise following model. The temperature rise following model generates a reference temperature rise trajectory corresponding to load changes based on real-time load current, and is corrected by combining ambient temperature and initial temperature conditions. The temperature rise following speed is limited by the effective thermal time constant. The temperature rise deviation sequence is the difference between the winding temperature and the reference temperature rise trajectory, and the recent sequence segments of the temperature rise deviation sequence are written into the deviation history queue field.

10. The intelligent fault diagnosis method for dry-type transformers based on the improved BGRU according to claim 9, characterized in that, After exiting the thermal response blind zone, the system performs separate diagnosis of overload conditions and winding short-circuit faults. This includes judging the changing trend of the temperature rise deviation sequence after the load current enters a stable or declining state. If the temperature rise deviation sequence decreases or tends to stabilize, it is judged as an overload condition. If the temperature rise deviation sequence continues to increase, it is judged as a winding short-circuit fault and a diagnosis alarm is triggered. Among them, the load current entering a stable or declining state includes the magnitude of the load current change rate being less than the preset stable threshold within a preset number of consecutive samplings or the load current showing a continuous downward trend.

11. A dry-type transformer fault intelligent diagnosis system based on an improved BGRU, used to implement the dry-type transformer fault intelligent diagnosis method based on an improved BGRU as described in any one of claims 1-10, characterized in that, include: The data acquisition and effective thermal time constant generation unit is used to acquire time series data, cooling fan start-stop status and thermal time constant, determine the cooling correction coefficient based on the cooling fan start-stop status, and calculate the effective thermal time constant based on the thermal time constant and the cooling correction coefficient. The mutation triggering and blind zone boundary calculation and marking unit is used to detect step or transient mutations in the load current and determine the mutation time. Starting from the mutation time, it determines the end time of the thermal response blind zone based on the effective thermal time constant and the preset proportional coefficient, and generates a thermal response blind zone marking sequence aligned with the time series data. The model inference and diagnostic control unit is used to input time series data into the improved BGRU model to output the original fault classification probability, and perform diagnostic control on the original fault classification probability based on the thermal response blind zone label sequence. When the maximum probability of the original fault classification probability corresponds to an overload state or a winding short circuit fault within the thermal response blind zone, it outputs an undetermined state and generates and locks the fault candidate cluster structure. The post-blind zone split diagnosis unit is used to split and diagnose the overload state and winding short circuit fault in the fault candidate cluster structure based on the changing trend of the temperature rise deviation sequence when the thermal response blind zone marking sequence indicates that it has exited the thermal response blind zone and the fault candidate cluster structure is in a locked state, and output the diagnosis conclusion. When the fault is diagnosed as a winding short circuit fault, a diagnosis alarm is triggered. The preset parameter storage area is used to store preset parameters for determining mutation triggering and the end time of thermal response blind zone, which are read by the mutation triggering and blind zone boundary calculation and marking unit.

Images